forked from huawei/mindspore2022
!29505 Fix AllGather Cast when the parameters are shared
Merge pull request !29505 from huangxinjing/fx_allgather_cast
This commit is contained in:
Gitee
commit
ea391f6eb1
|
|
@ -382,6 +382,10 @@ AnfNodePtr GetChildCastNode(const AnfNodePtr &node_ptr, const NodeUsersMap &node
|
|||
}
|
||||
auto users = node_users_map.at(node_ptr);
|
||||
for (auto &node_user : users) {
|
||||
cnode = node_user.first->cast<CNodePtr>();
|
||||
if (!cnode || !cnode->in_forward_flag()) {
|
||||
continue;
|
||||
}
|
||||
if (node_user.first) {
|
||||
visited.push(node_user.first);
|
||||
}
|
||||
|
|
@ -390,10 +394,10 @@ AnfNodePtr GetChildCastNode(const AnfNodePtr &node_ptr, const NodeUsersMap &node
|
|||
queue_node = visited.front();
|
||||
visited.pop();
|
||||
cnode = queue_node->cast<CNodePtr>();
|
||||
if (!cnode || !cnode->in_forward_flag()) {
|
||||
// MAKE_TUPLE will not appear after the load in the forward graph
|
||||
if (IsInNodeList(cnode, {MAKE_TUPLE})) {
|
||||
continue;
|
||||
}
|
||||
if (IsInAllGatherNodeList(cnode) || IsInNodeList(cnode, {LOAD, RESHAPE, DEPEND, UPDATESTATE, MAKE_TUPLE})) {
|
||||
} else if (IsInAllGatherNodeList(cnode) || IsInNodeList(cnode, {LOAD, RESHAPE})) {
|
||||
auto node_set = node_users_map.at(queue_node);
|
||||
for (auto &node_user : node_set) {
|
||||
visited.push(node_user.first);
|
||||
|
|
|
|||
|
|
@ -203,16 +203,16 @@ class Primitive(Primitive_):
|
|||
if not _is_in_auto_parallel_mode():
|
||||
if in_strategy is not None:
|
||||
logger.warning(f"The in_strategy of the operator in your network will not take effect in {mode} mode. "
|
||||
f"This means the the shard function called in the network is ignored. "
|
||||
f"This means the the shard function called in the network is ignored. \n"
|
||||
f"If you want to enable it, please use semi auto or auto parallel mode by "
|
||||
f"context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL "
|
||||
f"or context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL")
|
||||
f"or context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL)")
|
||||
if out_strategy is not None:
|
||||
logger.warning(f"The out_strategy of the operator in your network will not take effect in {mode} mode."
|
||||
f" This means the the shard function called in the network is ignored. "
|
||||
f" This means the the shard function called in the network is ignored. \n"
|
||||
f"If you want to enable it, please use semi auto or auto parallel mode by "
|
||||
f"context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL "
|
||||
f"or context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL")
|
||||
f"or context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL)")
|
||||
|
||||
self.add_prim_attr("in_strategy", in_strategy)
|
||||
self.add_prim_attr("out_strategy", out_strategy)
|
||||
|
|
|
|||
|
|
@ -0,0 +1,151 @@
|
|||
# Copyright 2022 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
""" test parallel optimizer shared test """
|
||||
import os
|
||||
import shutil
|
||||
import glob
|
||||
import numpy as np
|
||||
|
||||
import mindspore as ms
|
||||
import mindspore.nn as nn
|
||||
from mindspore import Tensor, Parameter
|
||||
from mindspore.common.api import _cell_graph_executor
|
||||
from mindspore.nn import TrainOneStepCell
|
||||
from mindspore.nn.wrap.cell_wrapper import _VirtualDatasetCell, MicroBatchInterleaved, PipelineCell
|
||||
from mindspore.nn.optim import Momentum
|
||||
from mindspore.ops import operations as P
|
||||
from mindspore import context
|
||||
|
||||
class Net(nn.Cell):
|
||||
"""Net definition"""
|
||||
def __init__(self, strategy1, strategy2):
|
||||
super(Net, self).__init__()
|
||||
self.fc1 = P.MatMul().shard(strategy1)
|
||||
self.fc2 = P.MatMul().shard(strategy2)
|
||||
self.p1 = Parameter(Tensor(np.ones([48, 64]).astype(np.float32)), name="weight1")
|
||||
self.p2 = Parameter(Tensor(np.ones([64, 16]).astype(np.float32)), name="weight2", parallel_optimizer=False)
|
||||
self.sub = P.Sub()
|
||||
|
||||
def construct(self, x, y):
|
||||
x = P.Cast()(x, ms.float16)
|
||||
p1 = P.Cast()(self.p1, ms.float16)
|
||||
p2 = P.Cast()(self.p2, ms.float16)
|
||||
x = self.fc1(x, p1)
|
||||
x = self.fc2(x, p2)
|
||||
return self.sub(x, y)
|
||||
|
||||
|
||||
class Net2(nn.Cell):
|
||||
"""Net definition"""
|
||||
def __init__(self, strategy1, strategy2):
|
||||
super(Net2, self).__init__()
|
||||
self.net1 = Net(strategy1, strategy2)
|
||||
self.net2 = Net(strategy1, strategy2)
|
||||
self.net1.pipeline_stage = 0
|
||||
self.net2.pipeline_stage = 1
|
||||
self.sub = P.Sub()
|
||||
|
||||
def construct(self, x, y):
|
||||
out1 = self.net1(x, y)
|
||||
out2 = self.net2(x, y)
|
||||
return self.sub(out1, out2)
|
||||
|
||||
|
||||
def auto_parallel_compile_net(mode, dev_num, net, strategy1=None, strategy2=None,
|
||||
interleaved_batch=2, stages=1, micro_size=1):
|
||||
context.set_context(mode=context.GRAPH_MODE)
|
||||
context.set_auto_parallel_context(parallel_mode=mode, device_num=dev_num, enable_parallel_optimizer=True,
|
||||
pipeline_stages=stages)
|
||||
inputs = Tensor(np.ones([64, 48]).astype(np.float32))
|
||||
label = Tensor(np.zeros([64, 16]).astype(np.float32))
|
||||
net = MicroBatchInterleaved(net(strategy1, strategy2), interleaved_batch)
|
||||
if stages > 1:
|
||||
net = PipelineCell(net, micro_size=micro_size)
|
||||
net = _VirtualDatasetCell(net)
|
||||
parameters = net.trainable_params() if stages == 1 else net.infer_param_pipeline_stage()
|
||||
optimizer = Momentum(parameters, learning_rate=0.1, momentum=0.9)
|
||||
train_network = TrainOneStepCell(net, optimizer).set_comm_fusion(4)
|
||||
train_network.set_auto_parallel()
|
||||
train_network.set_train()
|
||||
_cell_graph_executor.compile(train_network, inputs, label, phase="train", auto_parallel_mode=True)
|
||||
context.reset_auto_parallel_context()
|
||||
return train_network
|
||||
|
||||
|
||||
class TestSharedParameterCast:
|
||||
def setup_method(self):
|
||||
self.output_path = './graphs' + self.__str__()
|
||||
context.set_context(save_graphs=True,
|
||||
save_graphs_path=self.output_path)
|
||||
|
||||
def teardown_method(self):
|
||||
shutil.rmtree(self.output_path)
|
||||
|
||||
def cat_fp16_from_ir(self, target_count):
|
||||
"""
|
||||
This function will check the float16 count with the golden one.
|
||||
:param target_count: The gold float16 count in the Ir files
|
||||
"""
|
||||
# Find the step_parallel_end
|
||||
ir_files = glob.glob(os.path.join(self.output_path, 'rank_0', '*_validate*.ir'))
|
||||
assert len(ir_files) == 1
|
||||
appear_count = 0
|
||||
with open(ir_files[0], 'r') as fp:
|
||||
for line in fp:
|
||||
if 'Float16' in line:
|
||||
appear_count += 1
|
||||
assert appear_count == target_count
|
||||
|
||||
def test_optimizer_fp16(self):
|
||||
"""
|
||||
Feature: CastBeforeAllGather.
|
||||
Description: The order should be load, cast(from fp32 to fp16), AllGather.
|
||||
Expectation: the number of the float16 tensor is not equal to 27, 27 is obtained by manually checked graph.
|
||||
"""
|
||||
auto_parallel_compile_net("semi_auto_parallel", 8, Net, ((8, 1), (1, 1)), ((8, 1), (1, 1)),
|
||||
interleaved_batch=1)
|
||||
self.cat_fp16_from_ir(target_count=27)
|
||||
|
||||
def test_optimizer_fp16_micro_batch(self):
|
||||
"""
|
||||
Feature: CastBeforeAllGather with MicroBatchInterleave applied.
|
||||
Description: The order should be load, cast(from fp32 to fp16), AllGather.
|
||||
Expectation: the number of the float16 tensor is not equal to 41, 41 is obtained by manually checked graph.
|
||||
"""
|
||||
auto_parallel_compile_net("semi_auto_parallel", 8, Net, ((8, 1), (1, 1)), ((8, 1), (1, 1)),
|
||||
interleaved_batch=2)
|
||||
self.cat_fp16_from_ir(target_count=41)
|
||||
|
||||
def test_optimizer_fp16_pipeline(self):
|
||||
"""
|
||||
Feature: CastBeforeAllGather with PipeLine applied.
|
||||
Description: The order should be load, cast(from fp32 to fp16), AllGather.
|
||||
Expectation: the number of the float16 tensor is not equal to 27, 27 is obtained by manually checked graph.
|
||||
"""
|
||||
auto_parallel_compile_net("semi_auto_parallel", 8, Net, ((8, 1), (1, 1)), ((8, 1), (1, 1)),
|
||||
interleaved_batch=1,
|
||||
stages=1, micro_size=1)
|
||||
self.cat_fp16_from_ir(target_count=27)
|
||||
|
||||
def test_optimizer_fp16_pipeline_micro_batch(self):
|
||||
"""
|
||||
Feature: CastBeforeAllGather with MicroBatchInterleave and PipeLine applied.
|
||||
Description: The order should be load, cast(from fp32 to fp16), AllGather.
|
||||
Expectation: the number of the float16 tensor is not equal to 41, 41 is obtained by manually checked graph.
|
||||
"""
|
||||
auto_parallel_compile_net("semi_auto_parallel", 8, Net, ((8, 1), (1, 1)), ((8, 1), (1, 1)),
|
||||
interleaved_batch=2,
|
||||
stages=1, micro_size=1)
|
||||
self.cat_fp16_from_ir(target_count=41)
|
||||
Loading…
Reference in New Issue